Octofs

Apply insert/replace operations (max 50) to one file atomically. Targets are line ids ("12:a3") from view or edit output, verified before anything is written; a stale id fails with the current content. All targets refer to the original file and must not overlap. Put every edit you already know for a file into one call: separate edit calls to the same file, even parallel ones, invalidate each other's ids. The result diff carries fresh ids for follow-up edits — no re-view needed; removed lines, and the middle of a long added block, show as an id range, and a trailing `shift:` line says how later original line numbers moved. Insert anchors 0 (file start) and -1 (end) are plain integers. Content is raw text without id prefixes.

Copy a line range from one file and append it into another without retyping it — moving code between files, splitting modules. The source is left untouched; to move, follow with a batch_edit that removes the range.

Run a shell command: builds, tests, git, project CLIs. `sh -c` (`cmd /C` on Windows) in the current workdir, local machine only, no stdin or TTY — never run interactive commands or prefix `cd <workdir> &&` (`cd <other> && …` is fine one-off; `workdir` switches permanently; a remote ssh:// workdir disables shell). Use `view` to read, list and search files (cat/head/tail/grep/rg/find/ls except as a later pipe stage, `sed -i` and standalone `tee` anywhere, and echo/printf/cat redirects into the workdir are rejected). Local writes outside it (`cd /tmp/x && printf … > f`) run with a hint; `printf … | prog` is no write. Output is terminal-clean (ANSI and progress redraws stripped, repeated lines collapsed with a count; pipe through `od -c` or `xxd` for exact bytes); a non-zero exit returns as an error with the output. Large recognized test/build output is digested with its full output path appended. Verify narrow first: start with the smallest check that exercises your change — one test, target, package or sample input, in the mode it targets — then broaden only as far as the change can reach; a change everything depends on (a shared header, a build flag) ends with the full run. Repeat a full run only after a change that could alter its result. A command still running after ~10s moves to the background and returns a job resource showing a progress tail while running; on exit you receive the exit code and the same output as a foreground command. Its result then arrives on its own as a message. Do not poll it (`ps`, `tail`, sleep, viewing the linked resource, any other status check), re-run it or start a variant of it (other flags or filters). While it runs, keep working on anything that does not need its result (read or review related code, write or update tests, prepare the next edit); only when nothing like that is left, reply with a brief status and no tool call to wait. Resume dependent work when its completion notification arrives. Unrelated commands run concurrently…

File operations. create: new file only — fails if it exists, parent directories are made. str_replace: raw file text (real newlines, no `N:hh|` prefixes); old_text must match exactly once, or set replace_all: true — on multiple matches the error lists their line ids, on none the closest matches, for batch_edit. delete: remove a file (not a directory). undo_edit: revert the last edit to `path` (10 levels, delete included). Prefer batch_edit when you already hold line ids.

Read files, list directories, search content. Lines render as `N:hh|content`; copy `N:hh` verbatim — it is the line id edit tools target. Listings are recursive and gitignore-aware, one `path NL ~Nt` (lines, ~tokens) per file. An image file (png, jpg, gif, webp by extension) comes back as the image itself; `start`/`end` and `content` do not apply to it. Re-viewing a whole file may return only the hunks changed since your last view (or an unchanged marker); ranged reads always return the requested lines. Reuse returned content: read a complete block in one call, combine adjacent windows, never re-read overlapping ranges or narrow one just to pick edit targets; re-read only when the file may have changed or output was truncated.

Switch the working directory for later calls (`path`) or revert to the session root (`reset: true`). Every tool resolves relative paths against it; do not call it just to check the directory. A remote (ssh://) workdir disables `shell`.